Research on the Structure of Pediatric Epilepsy Electronic Medical Records Based on Transfer Learning

Medical institutions commonly utilize electronic medical records (EMRs) to document patients’ medical conditions, which contain valuable medical information. However, EMRs often consist of semi-structured or unstructured data, presenting significant challenges in processing and analysis. This paper addresses the needs of subsequent tasks such as assisting diagnosis and information extraction (IE). We present the process of structuring electronic health records based on pediatric epilepsy cases, including EMRs structural analysis, data preprocessing, named entity recognition (NER), and data integration. NER is a critical step in the processing pipeline. Therefore, we propose a pediatric epilepsy EMRs entity recognition model, T-RoBERTa-BiLSTM-CRF, based on transfer learning. After training on the constructed pediatric epilepsy dataset, T-RoBERTa-BiLSTM-CRF is used to perform NER tasks in EMRs structuring, achieving an F1 score of 79.25 \(\%\) . Identified medical entities are integrated at the text level to achieve EMRs structuring aligned with the needs of subsequent tasks.